This repository provides implementation and some experimental results of Direct Token Optimization. The repository suppports unlearning TOFU dataset and MUSE-HP dataset. To evaluate, please refer to MUSE Benchmark and open-unlearning.
Configuration files are located in the configs/ directory and can be customized for different experiments.
- Tofu dataset is available in Huggingface.
- For MUSE harry potter dataset, download the harry potter dataset from MUSE-Benchmark
The project requires Python 3.10+ with dependencies listed in requirements.txt, including PyTorch, Transformers, Hydra, and other ML libraries.
We provide an example script to run. Please refer to run_forget01.sh.
Under the results folder, we provide an evaluation result from TOFU benchmark.